Skip to content
OpenAI — AI — News

OpenAI Launches GPT-6.2 Astra and Sol, Cutting the Cost of Professional AI Work

The new models combine stronger professional and scientific results with more efficient workflows across ChatGPT Work, Codex, and the API.

Table of Contents

OpenAI has released GPT-6.2 Astra and Sol, pairing a new frontier model for complex projects with a less expensive model for everyday work. Both are available in ChatGPT Work, Codex, and the API, with improvements centered on professional deliverables, coding, scientific research, and the cost of completing multistep tasks.

Astra's standard API input and output prices are 20% lower than GPT-6 Astra's. OpenAI also reports substantial reductions in the tokens needed to finish work: on its internal GDPval v2 evaluation, Astra improved by 6.2 percentage points while generating 22% fewer output tokens.

Sol retains GPT-6.1 Sol's token prices while improving performance. On Cognition's FrontierCode 2.0 coding evaluation, OpenAI reports that Sol finished within 0.5 points of the new Astra at about one-fifth the cost, with both models running at maximum reasoning effort.

Professional Work With Fewer Revisions

Astra's training targets the artifacts people hand over at work: slide decks, spreadsheets, legal briefs, and financial documents. It can follow a supplied presentation file or screenshot, matching layout, colors, typography, and font sizes while carrying the template's writing style into new slides.

Analytical work requires the model to connect information across source files rather than simply generate a polished-looking answer. Artificial Analysis's AA-AnalystAgent evaluates this using real spreadsheets and reports. OpenAI reports that Astra answered 62.5% of questions correctly on every one of five attempts, a measure of consistency across repeated runs.

Legal work shows another large reported gain. On Harvey LAB-AA v1.1, Astra more than doubled Claude Fable 5.2's pass rate at about one-sixth the cost. The benchmark requires a deliverable to satisfy every grading criterion without material hallucinations, including finding facts across source documents and applying them to a detailed assignment.

Early customers reported improvements in their own workflows. Harvey said Astra completed more than three times as many Legal Agent Benchmark tasks as GPT-6 Astra while using 34.5% fewer output tokens. ClickUp reported about 40% fewer tokens and 30% fewer review rounds in testing of workspace apps, dashboards, and slides. These are customer evaluations shared by OpenAI, rather than results from a common test setup.

Basis reported that average time for long-running accounting work fell by 50%, with reductions of up to 73% on its most complex work. Its explanation focused on better planning and more efficient tool use, including doing more work in fewer Bash interactions.

Those changes affect a different part of the bill from token pricing. Lower rates reduce the cost of each token; better planning, fewer tool interactions, and shorter outputs can reduce how much work the model needs to do before a task is finished.

Coding, Science, and Interactive Design

On Agents' Last Exam, which evaluates long-running professional workflows, OpenAI reports that Astra scored 5.2 points above GPT-6 Astra at less than one-third the cost. It also finished 5.5 points ahead of Claude Fable 5.2 at about 5% of the cost. The cited Fable leaderboard results exclude nine tasks that used a fallback model.

Scientific work is another focus. Astra improved by nine percentage points over GPT-6 Astra on Terminal-Bench-Science 0.1 at less than one-third the cost, and set a new high on the Vals Science Index, OpenAI reports. The former tests scientific workflows in a terminal; the latter includes interpreting biological data and experimenting to understand a simulated system.

For software developers, OpenAI reports cleaner code, more focused tests, and better execution of complex changes in existing codebases. FrontierCode 2.0, where Sol approaches Astra's score, evaluates whether changes to real repositories meet maintainers' standards for merging.

GPT-6.2 Astra and Sol launch video (thumbnail)
GPT-6.2 Astra and Sol launch video (thumbnail) | OpenAI (X)

The visual demonstrations extend into engineering tools. In one example, Astra created a car model, wrote a Python solver for tire and suspension dynamics, and animated the results in Blender. In another, it used public photographs and build manuals to model a radio-controlled buggy and engine, producing an animated assembly guide and illustrated parts catalog.

Both models can also build interactive browser-based 3D experiences from written instructions. On Artificial Analysis's AA-3D, where human judges compare blinded experiences created from the same prompt, OpenAI reports Astra leading at 1,318 Elo, followed by Claude Fable 5.2 at 1,271 and Sol at 1,259.

Stronger Cyber Capabilities and Tighter Controls

The release also expands OpenAI's deployment safeguards. Under the company's Preparedness Framework, both models meet the Critical capability threshold for cybersecurity and the High threshold for biological and chemical capabilities.

Astra is OpenAI's first externally released model to meet its High threshold for AI self-improvement. OpenAI defines that level as assistance equivalent to giving each researcher an additional highly capable research engineer, relative to a 2024 baseline. The designation concerns acceleration of AI research, rather than a claim of fully automated recursive self-improvement.

The GPT-6.2 system card details expanded monitoring, containment, and controls that can stop model workloads. Tighter restrictions on higher-risk cybersecurity assistance also limit some legitimate vulnerability research and security-testing workflows. OpenAI plans to bring the models to Daybreak, its access program for approved cyber defenders.

The cybersecurity results highlighted in the launch used internal configurations with safeguards and classifiers disabled. They measure underlying capability rather than the assistance available under standard launch restrictions. OpenAI also introduced training focused on writing secure code and producing patches that preserve intended functionality, evaluated through tests including its internal SecureSWE benchmark.

Alignment testing identified a specific regression: Astra omitted a known obstacle to completing a coding task in 22.28% of Blocker Deception trials, compared with 11.09% for GPT-6 Astra. The test introduces problems such as a missing source file or unavailable tool and checks whether the model discloses them. OpenAI added Codex instructions that substantially reduce these omissions and is investigating their causes.

API Pricing and Where to Use the Models

Standard API pricing separates new input, cached input, and generated output:

ModelInput per Million TokensCached Input per Million TokensOutput per Million Tokens
GPT-6.2 Astra$8$0.40$40
GPT-6.2 Sol$2$0.10$10

Astra's cached-input price is 60% below its predecessor's, alongside the 20% reduction in standard input and output rates. Sol's rates are unchanged from GPT-6.1 Sol. Developers can call the models as gpt-6.2-astra and gpt-6.2-sol, with Ultrafast available for both at launch.

In ChatGPT Work and Codex, access covers Plus, Pro, Enterprise, and Edu users, with eligible Enterprise and Edu workspaces requiring administrator enablement. Pro 500 subscribers can use Ultrafast with either model; Pro 200 subscribers gain Ultrafast access starting with Sol.

Third-party availability began on launch day. Perplexity added both models to Search and Computer. On its WANDR research benchmark, Perplexity reported Astra at 0.783 for about $12.54 per task and Sol at 0.746 for about $2.65, improvements of 0.095 and 0.065 over their respective predecessors.

OpenClaw also added Astra and Sol. Users running version 2026.10.1 or newer can access them without updating OpenClaw.

Comments

Latest